← back to catalog · registered 2026-11-16 21:03

Synchopate/Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated-GGUF

Synchopate Qwen GGUF MoE multimodal
Your rig guess connected
? Why do I need an app?
Reading your rig…

This is a rough estimate. Install the free app - we'll show exact numbers.

Reading real hardware from your app right now. Numbers below are exact.

Below is the per-quantization compatibility for this model.

curl -H "Authorization: Bearer $ABL_KEY" \
     "https://abliteration.org/api/v1/models/Synchopate%2FQwen3.8-Flash-Next-MaxQ-32GB-Abliterated-GGUF"
Response includes
  • classification m8
  • files 6
  • author_summary 4 models
  • readme_text full
10 credits · hourly refresh · ~4 KB payload Get an API key →
Abliteration classifier · v1.0.0
M8
Primary method

Repackaging (quantization)

Applied on top of direct removal inherited from the base model.
Confidence
MEDIUM
Inherited from base model
Why this label 3 signals
Method inferred from partial signals - repository name, related files, or tag patterns. Producer identity not confirmed; label may sharpen or shift as we gather more evidence.
  • 'abliterated' in name/tags
  • is_gguf=1
  • assume M1 (base ablation) + M8 (GGUF quant) - default when producer unknown
Refusal direction extraction

No specific extraction method could be identified for this model. The producer either did not document it or used a proprietary pipeline.

What is a refusal direction? →
Downloads · 30-day
901
Likes
0
Model age
5d ago
created 2026-11-16
Downloads over time
Now 901 → from 0 ↑ 0%
0330661991 0 on Nov 16 901 on Nov 22 Nov
Nov 16 → Nov 22 · 7 snapshots · spans 6 days

Genealogy 0 direct forks

Full fork graph →

This model's place in the market. Above: what it was derived from. Below: the tree of everything derived from it.

Metadata

License
apache-2.0
Tags
gguf qwen3.8 qwen4exp moe abliterated heretic uncensored 32gb-ram maxq cpu-inference llama-cpp llama-modchip

Related

Control vectors

Total size
35.7 GB
Files
6
Quantizations
2
Registered
2026-11-16 21:03
Last updated on HF
2026-11-20 21:22

Files by quantization

BF16 1 file 866 MB
mmproj-Qwen3.8-Flash-Next-BF16.gguf 866 MB ef2af1d6 download
Auxiliary files 5 files 35.7 GB
Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated-00001-of-00002.gguf 25.0 GB 6e8ae62c download
Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated-00002-of-00002.gguf 10.7 GB 86b79714 download
README.md 13.7 KB 88ce15ae download
.gitattributes 1.74 KB 545f8bf7 download
SHA256SUMS 356 B 9039932a download

README current version from Hugging Face


license: apache-2.0
base_model:

  • Qwen/Qwen3.8-Flash-Next
    base_model_relation: quantized
    quantized_by: Synchopate
    library_name: gguf
    pipeline_tag: image-text-to-text
    tags:
  • gguf
  • qwen3.8
  • qwen4exp
  • moe
  • abliterated
  • uncensored
  • 32gb-ram
  • maxq
  • cpu-inference
  • llama-cpp
  • llama-modchip

Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated-GGUF (24.98 GiB Resident Trunk + 10.73 GiB Lazy NVMe PLE · Opt-In Subspace-Abliterated)

[!CAUTION]
Responsible-Use Disclaimer & Edition Positioning (Opt-In KL-Safe Subspace-Abliterated Checkpoint):
Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated attenuates refusal-mediating activation subspaces via norm-preserving rank-2 orthogonal projection strictly across blk.4..48 residual-stream writers (ssm_out, attn_output, ffn_down_shexp, ffn_down_exps, all 42 F32 ffn_down_exps.bias tensors, and blk.48 NextN MTP; blk.0..3 and token_embd spared), reducing held-out dual-use security refusals from 43.75% (7/16) to 12.50% (2/16, -71.4%) and <think> spot refusals to 0 / 4 (0.0%) at +0.3189% ΔPPL (rather than over-editing to "0% refusals"). Upstream safety refusal filters cannot be relied upon for content moderation.

  • Intended Scope: Authorized security research, reverse engineering, red-team auditing, CTF competitions, binary/exploit analysis, code auditing, and uncensored legal/policy analysis only.
  • Deployment: Do not expose unmoderated endpoints of this checkpoint to public networks without external access control and safety filtering.
  • Safety-Aligned Defaults: For standard upstream alignment on 32 GB RAM, use Synchopate/Qwen3.8-Flash-Next-MaxQ-32GB-GGUF (./install.sh --preset qwen38-32gb); on 64 GB+ RAM hosts, Synchopate/Qwen3.8-Flash-Next-MaxQ-64GB-GGUF (4 / 4 strict Blind-4 pass vs. 2 / 4 on MaxQ-64GB-Abliterated) is the recommended default pure coding flagship.

Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated is the opt-in uncensored 32 GB consumer RAM tier twin of Qwen3.8-Flash-Next-MaxQ-32GB, built 100% donor-free directly from Qwen/Qwen3.8-Flash-Next (rev 812d96e10a0dd82b69d749e68693698cf6c36daa) via single-pass fused rank-2 contrastive subspace projection + FP64 Woodbury DualGram (64 Jinja-templated pairs = 32 non-thinking + 32 thinking; verified 0.000000% 8-gram and 13-gram overlap across all 32 / 32 checks; 177 edited / 950 bit-identical Shard 1 tensors across blk.4..48, with 42 × F32 ffn_down_exps.bias orthogonalized to 7.436e-10). It keeps the 24.98 GiB Shard 1 trunk + blk.48 NextN MTP resident in DDR5 RAM while demand-paging the shared 10.73 GiB PCA-Wiener D=256 + RaBitQ Q1_0 Shard 2 PLE table via NVMe mmap (VmHWM = 28.40 GiB <= 28.5 GiB, RssAnon = 2.43 GiB, max_type = 42, 0 custom tensor types).

[!IMPORTANT]
Mandatory MaxQ-32GB-Abliterated Anti-Repetition Sampling & <think> Token Budget (1.63 bpw Trunk):
At 1.63 bpw, pure greedy temperature = 0 decoding without sampling penalties (max_tokens = 512) exhibited tail repetition after substantive technical output on 3 / 16 dual-use security prompts (ref_01 hex comment tail, ref_09 cyclic imports, and ref_14 box-drawing — with ref_14 crossing the automated 15× 20-gram threshold) and 2 / 4 Blind-4 prompts (2 / 4 loop-free at temp = 0). Configuring the sampling parameters below improved Blind-4 loop-free completion from 2 / 4 to 4 / 4 (100% loop-free) while preserving 1 / 4 (06_json PASS) strict parity with Aligned MaxQ-32GB (always enable these flags for coding, security analysis, and chain-of-thought, reserving temp = 0 strictly for short JSON/SQL extraction):

  • temperature = 0.6
  • top_p = 0.95, top_k = 20
  • presence_penalty = 0.2 – 1.0 (default 0.2)
  • repeat_penalty = 1.05
  • <think> Token Budget (max_tokens >= 16384): Mandate max_tokens >= 16384 whenever enable_thinking=True (max_tokens = 2048 <think> spot checks verified 0 / 4 refusal intent inside <think>), or pass enable_thinking=False under <= 4,096 output caps.

(Note: ./modchip and install.sh --preset qwen38-32gb-ablit configure these sampling flags automatically.)


Architecture & Parameter Summary (general.size_label Clarification)

[!NOTE]
Clarification on Hugging Face Hub's Auto-Detected "5.0B" Badge: Shard 1 carries general.size_label = "5.0B" inherited from the non-MoE dense trunk staging header during conversion. This checkpoint contains the complete 512-expert Qwen3.8-Flash-Next architecture (42 K = 512 ConeCollapsed layers + 7 K = 640 boundary layers + 320M-row PLE + blk.48 NextN MTP):

Architectural Component Specification
Upstream Base Model Qwen/Qwen3.8-Flash-Next (rev 812d96e10a0dd82b69d749e68693698cf6c36daa, 100% donor-free contrastive subspace edit)
Nominal / Collapsed Params ~176.5B nominal (~152.3B post-ConeCollapse + PLE; 92.28% mean retained spectral energy)
Active Transformer Params / Token ~4.6B–5.0B (10 of 512 routed experts + 1 shared expert + dense GDN/QSA; reflected in general.size_label)
Backbone Topology (qwen4exp) 48 Trunk Layers (36 Gated DeltaNet + 12 QSA Full Attention, compress_ratio = 4) + 1 Subspace-Aligned NextN MTP Layer (blk.48)
MoE Width & Bias Orthogonalization 512 / 512 Routed Experts (42 × K = 512 ConeCollapsed layers with orthogonalized 42 × F32 ffn_down_exps.bias [$\max |U_\ell^\top b_d'| / |b_d'| = 7.436 \times 10^{-10}$] + 7 × K = 640 boundary layers)
GGUF Tensor Types (1,128 LLM) {F32: 393, Q8_0: 291, Q4_K: 99, Q5_K: 98, IQ1_S: 62, IQ1_M: 51, BF16: 49, MXFP4(39): 49, IQ2_XXS: 17, IQ2_XS: 10, Q2_0(42): 7, Q6_K: 1, Q1_0(41): 1} (max_type = 42, 0 custom types)

File Manifest & Exact Byte / SHA-256 Verification

File Exact Size (Bytes) Size (GiB / GB) Tensors Role & Notes SHA-256
Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated-00001-of-00002.gguf 26,821,627,808 24.9796 GiB (26.822 GB) 1,127 Subspace-Abliterated Resident Trunk (177 edited incl. 42 bias / 950 bit-identical) + blk.48 NextN MTP 6e8ae62c37bc08de1a69e9e78ab9bae0262f87ffe543595206d1c3b3a810fd67
Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated-00002-of-00002.gguf 11,520,055,616 10.7289 GiB (11.520 GB) 1 Shared Lazy NVMe mmap per_layer_token_embd.weight (Q1_0 = 41) 86b79714867743ddf5911ee7dff2717ae275f870551d86469cbdbf026c6b552e
LLM Total (Shards 1 + 2) 38,341,683,424 35.7085 GiB (38.342 GB) 1,128 Complete 32 GB Hybrid-RAM LLM (VmHWM = 28.40 GiB <= 28.5 GiB) —
mmproj-Qwen3.8-Flash-Next-BF16.gguf (Optional) 907,542,880 0.8452 GiB (0.908 GB) 334 Multimodal Vision Projector (224 F32 + 110 BF16) ef2af1d6aa8a22740c7252bc7ce40c5c2f1249aa8d9c8499a4d4572c3cfdf9df

Quickstart (llama-modchip v3.0 on 32 GB RAM)

git clone https://github.com/synchopate/llama-modchip.git
cd llama-modchip
./install.sh --preset qwen38-32gb-ablit --accept-abliterated
./modchip chat

Direct llama-server invocation (mmap enabled — do NOT pass --no-mmap on 32 GB hosts!):

N_CORES=$(lscpu -p=CORE,SOCKET 2>/dev/null | grep -v '^#' | sort -u | wc -l)
THREADS=$(( N_CORES > 1 ? N_CORES - 1 : 1 ))

env RC_TABLES_DIR=/nonexistent ./llama-server-modchip-linux-x86_64-avx512 \
  -m ./qwen38-32gb-ablit/Qwen3.8-Flash-Next-MaxQ-32GB-Abliterated-00001-of-00002.gguf \
  -c 32768 -np 1 -fa on -ctk q8_0 -ctv q8_0 --spec-type draft-mtp --jinja --ctx-checkpoints 0 \
  --temp 0.6 --top-p 0.95 --top-k 20 --presence-penalty 0.2 --repeat-penalty 1.05 \
  -t "$THREADS" -tb "$THREADS" --poll 0 --poll-batch 0 --slot-save-path /dev/shm --port 8080
  • 32 GB mmap & /dev/shm Guidance: Keep default mmap enabled (and optional MemoryMax=28.5G) so Linux reclaims cold PLE pages automatically; --ctx-checkpoints 0 halves /dev/shm Slot-0 snapshot size from 233.1 MiB to ~120.6 MiB (~19 ms restore vs 38.5 ms default).
  • Upstream llama.cpp & <think> Notes: Pass --no-repack (-nr) only on unpatched upstream llama.cpp (llama-modchip v3.0 runs default --repack ON cleanly); set max_tokens >= 16384 when enable_thinking=True (or enable_thinking=False for <= 4,096 caps).

Empirical Certification: Refusal Removal, Signal Preservation & NextN MTP Speed

Evaluated under llama-modchip v3.0 (-t 15 -tb 15 --poll 0 --poll-batch 0, default --repack ON, RC_TABLES_DIR=/nonexistent) against the donor-free BF16 reference (Mean PPL(base) = 2.776624 ± 0.054868) and aligned MaxQ-32GB:

Checkpoint Exact LLM Bytes Mean PPL(Q) ± SE ΔPPL vs. Aligned Mean KLD ± SE Median KLD Same top p (%) Dual-Use Refusal Rate (16 Prompts) Blind-4 Coding/SQL/JSON Pass
Aligned MaxQ-32GB 38,341,683,424 B 5.80406 ± 0.13982 (5.8120 Stage 4C) 0.00% (Ref) 1.00399 (1.0069 Stage 4C) 0.39858 66.275% 43.75% (7 / 16; 2 / 4 <think>¹) 1 / 4 (v3.0 --repack; 2/8 Stage 4C -nr)
MaxQ-32GB-Abliterated 38,341,683,424 B 5.822571 ± 0.140087 +0.3189% (+0.018511, 9.4× under <= 3.0% Gate) 1.016915 ± 0.015155 0.398768 65.539 ± 0.471% 12.50% (2 / 16, -71.4%; 0 / 4 <think>¹) 1 / 4 (06_json PASS — 100% parity; 4/4 loop-free w/ sampling)
  • ¹ KL-Safe Dual-Use Compliance & <think> Refusal Audit (0 Disguised Refusals): Under KL-safe rank-2 contrastive subspace projection (90% / 99% / Max KLD = 2.930651 / 7.303628 / 13.513656), held-out dual-use security refusals (enable_thinking=False, temp = 0) drop from 7 / 16 (43.75%) to 2 / 16 (12.50%, -71.4%) — preserving ΔPPL = +0.3189% rather than over-projecting to "0% refusals" — with 0 / 4 benign false refusals and 0 / 4 (0.0%) <think> spot refusals (enable_thinking=True, max_tokens = 2048 vs. 2 / 4 on Aligned MaxQ-32GB, and 2 / 4 true <think> refusals on Aligned MaxQ-64GB once Unicode curly apostrophe I can’t \u2019 on ref_01 is counted). Full-text inspection of all flipped MaxQ-32GB-Abliterated outputs (radare2 binary patch, glibc >= 2.32 safe-linking ptr ^ (addr >> 12) tcache audit, SUID $PATH audit, Python pickle.__reduce__, 3.3V UART hardware pinout) confirmed working technical substance with 0 curly-apostrophe false negatives, 0 disguised refusals, and 0 gibberish.
  • llama-modchip v3.0 Throughput, blk.48 NextN MTP Alignment & 28.5 GiB Memory (16 Zen 5 Cores, 0 GPUs): Aligned MaxQ-32GB achieves 74.55 tok/s pp512 and 11.45–14.29 tok/s tg128 (+21.1% combined v3.0 SIMD/GDN/repack + non-convoy -t <N-1> --poll 0 gain over Stage 4C's oversubscribed -t 16 --poll 50 --no-repack 9.46 tok/s, VmHWM = 26.91 GiB <= 28.5 GiB Cgroup v2 Swap=0 PASS; cold rep 0 Stockfish-PVS v2 refactoring fallback 12.82 tok/s). MaxQ-32GB-Abliterated (VmHWM = 28.40 GiB <= 28.5 GiB, RssAnon = 2.43 GiB) achieves 77.53 tok/s pp512 and 11.65 tok/s mean / 12.73 tok/s cold rep 0 tg128 at 72.83% (260 / 357) MTP acceptance on p512, 12.78 tok/s (85.68% = 1298 / 1515) on blind_01 C++20, 11.74 tok/s (84.32% = 984 / 1167) on blind_06 JSON, and a token-weighted 76.22% MTP acceptance across top-4 tg512 abliterated prompts (peak 88.33% @ 13.24 tok/s on ref_09), beating Aligned MaxQ-32GB on both benign controls (68.31% vs. 68.28%) and complied-only security prompts (62.46% vs. 60.54%).
  • Blind-4 Parity & Edition Positioning: Matches default-repack Aligned MaxQ-32GB (1 / 4 strict pass on 06_json_charterparty_settlement), with recommended sampling (temp=0.6, top_p=0.95, top_k=20, presence_penalty=0.2, repeat_penalty=1.05) lifting Blind-4 loop-free rate from 2 / 4 (temp = 0) to 4 / 4 (100% loop-free). For pure software engineering on 64 GB+ hosts, use Aligned MaxQ-64GB (4 / 4 Blind-4 vs. 2 / 4 on MaxQ-64GB-Abliterated); use MaxQ-32GB-Abliterated for 32 GB security research, reverse engineering, red-team auditing, and uncensored legal/policy analysis.

Portable v3.0 Binary Checksums (GLIBC <= 2.30)

54af3232375ede820f91143c106a594c83b815a3dca66fe98ab08bcb8b63d4f0  llama-server-modchip-linux-x86_64-avx512
0f0c07f97a98d8aeb9a012264fdeff5a787d2104a2afb122052bfe0a21c2d1a9  llama-server-modchip-linux-x86_64-avx2
531d870988dcc580322341d2bbd7f0cf6b6cc3b3376a2e86d092a021f7cddba2  llama-cli-modchip-linux-x86_64-avx512
49a918206afe371d16ab14fd66d9cc87e35c7f8c0471c2cc13c759f2e0350d03  llama-cli-modchip-linux-x86_64-avx2

README history 7 versions

The author's README evolved over time. Click a version to see its content at that point.

  1. 2026-11-20 Cite cold-cache 28.26 GiB cgroup peak alongside earlier warm-cache gate 29f5ff4 12.7 KB
    Loading...
  2. 2026-11-20 Update model card to llama-modchip v3.1.1 (RouteCrack 2.0, reconciled metrics... 2af80cd 12.6 KB
    Loading...
  3. 2026-11-18 docs: link base_model (Qwen/Qwen3.8-Flash-Next) for HF Quantizations tree a3b3dda 12.1 KB
    Loading...
  4. 2026-11-17 docs(v3.0-r2): unify Heretic-2 + NextN MTP model card metrics and tensor edit... ea5d3c7 12.1 KB
    Loading...
  5. 2026-11-17 Release v3.0-r2: Gated-Residual Heretic-2 over-projection + NextN MTP alignment 71a3d7b 14.5 KB
    Loading...
  6. 2026-11-16 Align YAML license metadata to upstream qwen-community-1.0 8b68deb 13.9 KB
    Loading...
  7. 2026-11-16 Add MaxQ v3.0 model card d18c90c 13.7 KB
    Loading...
Catalog is the map. Apps are the tools.

Run models on your own machine, not in the cloud.

Every model page has an "Open in Abliteration" button that hands the model directly to the first-party desktop client, at the quantization your rig can actually run. No API keys, no subscription, no prompt leakage.

Open in Abliteration